Is this only an issue if you are using agents? Or are there any theoretical risks of serving a model over say llama server (llama.cpp) and accessing it for chat in browser on machines on your local network?